DocChat

Tech Stack

Role

Full-stack Developer

Team Size

1

Duration

February 2026 - Ongoing

Project Overview

DocChat is an AI-powered API documentation assistant designed to help developers locate endpoints, understand parameters, and explore data structures through natural language. The system uses a deterministic-first + LLM-fallback two-phase routing architecture — ~80% of queries are resolved without any LLM calls, significantly reducing costs. Currently launched with Opta/Stats Perform sports data APIs as the first use case, supporting Motorsport (7 feeds) and Soccer (92 feeds) with a 4-layer knowledge injection system for precise answers. The long-term vision is a general-purpose API documentation platform where users can upload any OpenAPI spec and auto-generate knowledge bases. The frontend features a terminal-style dark theme (neon green + Matrix rain background), SSE streaming conversations, session history management, and user authentication.

Highlights

• Two-phase routing — deterministic keyword matching first (~2650 triggers), LLM fallback, ~80% queries with zero LLM calls
• 4-layer knowledge injection — endpoint-level docs → domain overview → shared knowledge base → dynamic topic routing (INDEX.yaml driven)
• Current use case: Opta sports data API — 99 endpoints covered (Motorsport 7 + Soccer 92), each with full knowledge files
• Terminal-style UI — neon green (#DCFF71) + Matrix rain background, Hack font, glassmorphism input
• SSE streaming chat + Streamdown Markdown rendering + syntax highlighting + bilingual (EN/CN) support
• Full user system — JWT auth (Access + Refresh tokens), conversation history, auto-generated session titles

Challenges

• Knowledge organization and precise routing across large-scale API endpoints, preventing LLM hallucination in massive documentation
• Data fusion from three sources: OpenAPI specs + PDF documentation + hand-written knowledge files
• SSE streaming communication with static-exported frontend (output: export)
• Docker containerized deployment with PostgreSQL network topology configuration

Solutions

• Built enhanced index JSON (Swagger parsing + PDF extraction + metadata injection), loaded into memory at startup
• Designed META.yaml with triggers + fields + scenarios for multi-dimensional matching, covering the vast majority of queries deterministically
• Vercel AI SDK v6 + TextStreamChatTransport for frontend SSE consumption
• Docker Compose with external network sharing PostgreSQL container, Nginx reverse proxy + Let's Encrypt HTTPS